Chinese startup DeepSeek has announced that it will make the code for its models available to all users next week, reaffirming its commitment to open source AI technology.
Image source: Solen Feyissa/unsplash.com
The company announced on social media platform X that it would open source five repositories, calling it “small but genuine progress” that it was sharing “with full transparency.” “These humble building blocks in our online service have been documented, deployed, and field-tested in a production environment,” DeepSeek said in its post.
DeepSeek last month released an open-source model of its DeepSeek R1 thinking machine that rivals US companies’ AI systems in performance, despite costing far less to build. That has made investors more critical of AI developers’ reports of the costs of deploying new models.
DeepSeek’s commitment to open source sets it apart from most AI firms in China, which, like their American rivals, prefer to market closed-source models, Reuters reports.
DeepSeek founder Liang Wenfeng told Chinese media in July last year that the company does not prioritize commercializing its AI models, and that open source can be a form of “soft power.” “When others follow your innovations, it gives you a deep sense of accomplishment,” Liang said.
The open source repositories will provide the infrastructure to support the AI models that DeepSeek has already shared publicly, building on existing open source model frameworks.
Earlier this week, DeepSeek introduced a new algorithm, Native Sparse Attention (NSA), designed to improve the training and inference performance of AI models in rich context.
Chatbot DeepSeek is the most popular in China with 22.2 million daily active users as of Jan. 11, according to Aicpb.com, surpassing the 16.95 million users of Chinese chat platform Douban.